Unbiased Classification Through Bias-Contrastive and Bias-Balanced Learning (NeurIPS 2021)

Official Pytorch implementation of Unbiased Classification Through Bias-Contrastive and Bias-Balanced Learning (NeurIPS 2021)

Setup

This setting requires CUDA 11. However, you can still use your own environment by installing requirements including PyTorch and Torchvision.

1. Install conda environment and activate it

conda env create -f environment.yml
conda activate biascon

2. Prepare dataset.

• Biased MNIST
  By default, we set download=True for convenience.
  Thus, you only have to make the empty dataset directory with mkdir -p data/biased_mnist and run the code.

• CelebA
  Download CelebA dataset under data/celeba

• UTKFace
  Download UTKFace dataset under data/utk_face

• ImageNet & ImageNet-A
  We use ILSVRC 2015 ImageNet dataset.
  Download ImageNet under ./data/imagenet and ImageNet-A under ./data/imagenet-a

Biased MNIST (w/ bias labels)

We use correlation {0.999, 0.997, 0.995, 0.99, 0.95, 0.9}.

Bias-contrastive loss (BiasCon)

python train_biased_mnist_bc.py --corr 0.999 --seed 1

Bias-balancing loss (BiasBal)

python train_biased_mnist_bb.py --corr 0.999 --seed 1

Joint use of BiasCon and BiasBal losses (BC+BB)

python train_biased_mnist_bc.py --bb 1 --corr 0.999 --seed 1

CelebA

We assess CelebA dataset with target attributes of HeavyMakeup (--task makeup) and Blonde (--task blonde).

Bias-contrastive loss (BiasCon)

python train_celeba_bc.py --task makeup --seed 1

Bias-balancing loss (BiasBal)

python train_celeba_bb.py --task makeup --seed 1

Joint use of BiasCon and BiasBal losses (BC+BB)

python train_celeba_bc.py --bb 1 --task makeup --seed 1

UTKFace

We assess UTKFace dataset biased toward Race (--task race) and Age (--task age) attributes.

Bias-contrastive loss (BiasCon)

python train_utk_face_bc.py --task race --seed 1

Bias-balancing loss (BiasBal)

python train_utk_face_bb.py --task race --seed 1

Joint use of BiasCon and BiasBal losses (BC+BB)

python train_utk_face_bc.py --bb 1 --task race --seed 1

Biased MNIST (w/o bias labels)

We use correlation {0.999, 0.997, 0.995, 0.99, 0.95, 0.9}.

Soft Bias-contrastive loss (SoftCon)

1. Train a bias-capturing model and get bias features.

python get_biased_mnist_bias_features.py --corr 0.999 --seed 1

2. Train a model with bias features.

python train_biased_mnist_softcon.py --corr 0.999 --seed 1

ImageNet

We use texture cluster information from ReBias (Bahng et al., 2020).

Soft Bias-contrastive loss (SoftCon)

1. Train a bias-capturing model and get bias features.

python get_imagenet_bias_features.py --seed 1

2. Train a model with bias features.

python train_imagenet_softcon.py --seed 1